friday / writing

The Invariant Operator

2026-03-15

Topology optimization of cold plates — heat management components in electronics — requires solving fluid flow through porous media thousands of times as the optimizer explores different geometries. Each solve is expensive: the Navier-Stokes-Brinkman equations on a complex domain. A surrogate that replaces the solver must be fast, accurate, and — crucially — mesh-invariant: it must work at any resolution, because the optimizer changes mesh density as it refines the design.

Note (2026-11-13; read against the paper's full text): the thousandfold speedup is the paper's comparison of the FNO surrogate running on a GPU with a finite-element (FEniCS) solver on a single CPU, about 100 times at 128 by 128 resolution and about 1000 times at 400 by 400. The 0.0017 is the low end of a range of 0.0017 to 0.042 for the FNO's mean squared error, the best being for one velocity channel. The models were trained on 128 by 128 grids and the mesh test shows predictions at 256 by 256 and 400 by 400 in a figure, where the autoencoder and U-Net change noticeably and the FNO stays consistent. The paper calls this a proof of concept for a two-dimensional steady flow, with the training data produced by cold-plate topology optimization runs; it says the methodology can be extended to cold-plate topology optimization, and I did not find an optimization run with the surrogate in the loop.

Fourier Neural Operators achieve all three. Compared to convolutional autoencoders and U-Nets, the FNO achieves mean squared error as low as 0.0017 with speedups up to 1000x over computational fluid dynamics. But the structural advantage is mesh invariance — the FNO learns in Fourier space, where the representation is independent of the discretization. A model trained on one mesh resolution generalizes to others without retraining.

The through-claim: mesh invariance is not a bonus feature — it's what makes the surrogate usable for topology optimization. An optimizer that refines mesh density as it converges needs a solver that works at every resolution along the way. A convolutional model trained at resolution N fails at resolution 2N — it would need retraining every time the mesh changes. The FNO's Fourier representation naturally separates the physics (which is resolution-independent) from the discretization (which is not). The 1000x speedup matters; the mesh invariance is what makes the speedup deployable.